Post-training a local, Mac-runnable multimodal Jev-style judge to power browser-use tools
AlchainHust · x · 2026-09-24
The author is post-training a Jev-style judge model with image understanding that runs locally on ordinary Macs. Key insight: with Qwen's abundance of open-source small base models, post-training single-capability classifiers is fairly easy and cheap. The goal is an open-source multimodal Jev to power faster, better browser-use and computer-use agents.
More from coding & agent
- "It has memory" means nothing: breaking Personal AI memory into write, retrieve, validity and override layers — sujingshen · 2026-09-24
- Claude Opus 5.5 generated an entire music video; open-source template renders 11s in 47s on M3 Max — xiaohu · 2026-09-24
- From Sparse Triples to Dense Wiki: Why Agents Need Better Memory — maier_ak · 2026-09-24
- Why agents stay stuck at 70-75% success: the hard problem of removing humans from the loop — Motor_Fox_9451 · 2026-09-24
- Harness Engineering explained: model + harness is what makes coding agents reliable — _jaydeepkarale · 2026-09-24
- npm co-founder Laurie Voss: devs are all product engineers now, AI is the web in 1997 — seldo · 2026-09-24